VLA-Stem: a "brain trunk" for robots doubles the success rate without collision of vision-language-action models

Rather than retrain a VLA model to avoid obstacles, a small policy learned by strengthening adjusts its movements on the fly: 52.8% successful tasks without collision, compared to 23.7% for the best alternative.

Vision-language-action (VLA) models can link a directive to gestures, but they are mainly designed to manipulate objects in an open space, regardless of spatial constraints or the obstacles that move. To refine them on avoidance data is costly to collect, especially for dynamic scenes, and may degrade their original skills. A Chinese team led by Bojun Zou and Jianmin Ji proposes another way with VLA-Stem.

Inspired by the brain stem, this light mechanism leaves the VLA model intact. It transforms the sequence of actions it produces into reference trajectories, then a policy driven by reinforcement, which also receives the robot's internal sensors and related obstacles, adjusts the execution in real time: it retains the intention of the task while changing the path and pace. Training combines teacher-pupil distillation and progressive learning.

In six simulated stress environments, VLA-Stem achieved 52.8% average success without collision, compared with 23.7% for the best competing safety mechanism, without loss of performance when the track is clear. Real robot tests also show fewer collisions against fixed obstacles as mobile.